Xcode 27 Coding Agents: From Autocomplete to Running Your App

Apple shipped Xcode 27 beta on June 8 at WWDC 2026 — and the Xcode 27 coding agents in this release represent the biggest shift in the iOS development workflow in years. Not because they write better code, but because they can now run your app and observe what happens.

From autocomplete to an autonomous loop

The jump is qualitative. Where the previous generation of Xcode AI could suggest a fix, Xcode 27 coding agents can execute a plan end-to-end: boot the simulator, build and install the app, synthesize touch events, capture screenshots, read diagnostic output, and iterate. The agent closes the loop the developer used to close manually.

The architecture that makes this safe is a planning layer. Before the agent touches anything, it produces an editable Markdown artifact describing exactly what it intends to do. You review, annotate, and approve the plan — then it executes. This distinction matters: you’re not watching an agent improvise, you’re approving a plan that an agent then carries out.

What Xcode 27 coding agents can actually do

The multi-tool surface in Xcode 27 is specific and worth mapping:

Simulator control: Boot, install, launch, and interact with running apps
UI interaction: Synthesize tap, swipe, and gesture events
Screenshot capture: Visual feedback that closes the observe-reason-act loop
Debug state: Pause and resume execution, evaluate expressions, interact with the LLDB console
Build configuration: Inspect and modify build settings, compiler flags, entitlements, and Info.plist
Diagnostics: Read crash logs, disk write warnings, energy reports, and hang data from Instruments

An agent that can run your app under test conditions, observe a failure, trace it to a build setting, and propose a targeted fix is a different category of tool from one that only reads source files. The first is a pair programmer. The second is an intelligent search.

MCP servers come to Xcode

The extensibility story is the one to watch long-term. Xcode 27 ships with full support for MCP (Model Context Protocol) servers inside coding agents. You can point Xcode at any MCP server and give the agent domain-specific tools: a design system lookup, an internal API client, a database schema reader, a deployment hook.

Combined with slash command plugins and a full ACP (Agent Client Protocol) implementation, Xcode’s coding intelligence becomes an open platform rather than a closed feature. MCP is the same protocol used by Claude and other frontier AI systems to give agents access to external tools. Apple shipping native ACP support means the broader agent ecosystem — existing servers, patterns, best practices — applies directly to Xcode workflows.

Teams that build internal MCP servers will get leverage from this immediately. A server that knows your API contracts, your component library, and your feature flags gives every developer on the team an AI with institutional memory.

Foundation Models gets a profiler

On the on-device AI side, Xcode 27 ships a dedicated Foundation Models Instrument. You can now trace Foundation Models calls end-to-end: inspect the exact prompt and system instructions sent to the on-device model, monitor response content and token counts, and measure inference latency directly in Instruments.

For apps already using the Foundation Models framework, this closes a significant observability gap. Previously, debugging on-device inference meant guessing what the model received. Now it’s traceable the same way a network call or memory allocation is traceable.

Apple also announced expanded developer access to Foundation Models at the WWDC 2026 Platforms State of the Union, with free access to the framework as the headlined announcement. The combination of broader access plus proper debuggability makes on-device inference production-viable in a way it wasn’t in the previous SDK cycle.

The hardware split you need to plan around

Apple’s new AFM 3 Core Advanced model — 20 billion parameters, running fully on-device — is exclusive to the three newest iPhone models. Every Foundation Models feature that requires AFM Core Advanced will be unavailable on the rest of the iOS install base.

TechRadar’s breakdown covers which devices qualify. The implication for product decisions: any feature gated to the most capable on-device model will reach only a fraction of users for the next two to three years. The Foundation Models framework provides graceful fallback tiers, but that requires explicit design — not an afterthought.

One more detail in the Xcode 27 changelog worth noting: Google Gemini is now available as a coding assistant option alongside existing providers. This tracks with Apple’s broader Google partnership announced at WWDC 2026, which includes Siri being rebuilt with Google AI. For developers, it means the underlying model powering Xcode’s agent is becoming a configuration choice.

What this means for the apps we build

The development process is where this lands practically. The simulator integration means Xcode’s AI can validate feature behavior, not just code structure — which is a step toward catching regressions before human review. That’s a real reduction in the cost of iteration, particularly on UI-heavy flows.

The MCP extensibility matters more as a multiplier. If your product has a design system, internal APIs, or domain-specific rules, encoding that once as an MCP server means every developer gets an AI that already understands your context. That’s where the compounding value lives.

We’re watching the AFM 3 hardware split carefully when evaluating on-device AI for client projects. If you’re deciding whether on-device inference is the right architecture for a feature, we’re happy to work through the trade-offs.

Sources

Xcode 27 Beta Release Notes — Apple Developer Documentation, June 8, 2026
Apple Developer Releases — Apple Developer, June 8, 2026
Apple Outlines Major AI and Developer Tool Updates at 2026 Platforms State of the Union — MacRumors, June 9, 2026
Apple unveils AFM 3 Core Advanced with 20 billion parameters for on-device AI at WWDC26 — Crypto Briefing, June 9, 2026
Only 3 iPhones can access the best version of Siri AI — TechRadar, June 10, 2026